ISTS42 Program/Agenda

The Lampedusa Sea Turtle Rescue Center, which has been in operation since 1990, has established strong relationships with the local fishery fleet in the archipelago, allowing the rescue of over 6,000 loggerhead sea turtles.

In this study, we analyzed data from 3,275 hospitalized Caretta caretta turtles up to the year 2023. Our primary goal was to identify factors that could impact the success of rehabilitating injured sea turtles at our facility. We thoroughly examined several key factors, including the type of clinical case, the health condition of the animals, the type of therapy administered and the presence of qualified surgeons.

For the first factor, we assessed the survival percentage among animals with various clinical cases, such as infections, flipper/carapace/head fractures or wounds, and the presence of hooks or fishing line entanglements in the digestive tract. Among these clinical cases, a few displayed a rehabilitation success rate below 50%, particularly cases involving head fractures and those with entanglements in various locations.

For the second factor, we evaluated therapy outcomes (recovery or death) and the health condition of the turtles (good health, depressed, or comatose) using the Fisher exact test. The results from the Fisher's test confirmed the significant influence of the turtles' health condition on the success of sea turtle rehabilitation (Fisher test=369.894; d.f.=2; P<0.001).

Regarding the last factor, we divided the study period into five subperiods (2001-2005, 2006-2010, 2011-2015, 2016-2020, 2021-2023) based on the evolution of surgical techniques and the presence of expert surgeons with direct experience in sea turtle surgery. Our ANOVA analysis further supported the significance of the professionals' experience (univariate ANOVA=4.953; d.f.=4; P=0.016). It became evident that bycatch and the health condition of the turtles significantly influenced the success of rehabilitation, while the presence of competent surgeons led to a substantial increase in sea turtle survival, aligning with our expectations.

Assessment of the respiratory system in sea turtles is crucial due to their unique anatomy, with the lungs situated beneath the carapace, making them susceptible to injuries, particularly from vessel strikes. These injuries can result in open lung wounds, leading to aspiration, loss of buoyancy control, and secondary infections. Pulmonary diseases in sea turtles can stem from various causes, including entanglement in marine debris such as plastic waste, fishing lines, and ghost nets, which can directly harm the airways or lead to severe secondary infections. This study aimed to outline the diagnostic methods and treatments for pulmonary diseases in sea turtles. It involved 40 turtles, all of which exhibited radiographic signs of pulmonary pathology. Among them, four underwent more extensive diagnostic procedures, including bronchoalveolar lavage, which allowed for fluid retrieval from the lower airways. This fluid was then subjected to cytological and bacteriological examinations in all 40 subjects. The findings indicated that radiographic examination proved instrumental in diagnosing pulmonary disorders in sea turtles. Bacterial cultures predominantly showed gram-negative strains, with a high level of antibiotic resistance, particularly against beta-lactams, Colistin sulfate, and Tetracycline. Treatment involved specific antibiotic therapies, such as Enrofloxacin and Ceftazidime, for a portion of the sea turtles. However, in cases where antibiotic resistance was prevalent against all antibiotics tested, lung disease resolution was achieved through techniques like coupage and environmental management, without antibiotic intervention. Ultimately, the study emphasizes the significance of thorough diagnostic procedures to attain accurate and early diagnoses, preventing unnecessary treatments and addressing antibiotic resistance in sea turtles. It underscores the importance of radiographic examinations as a primary screening tool for turtles displaying respiratory symptoms or abnormal buoyancy. Additionally, susceptibility testing with antimicrobials played a pivotal role in tailoring appropriate therapies, and contributing to the reduction of antibiotic resistance. All 40 sea turtles involved in the study were successfully released back into the sea following the treatments.

Sea turtle species are important for preserving marine ecosystems and the maintenance of their biodiversity. The conservation of these species is carried out through scientific studies focused on population monitoring, migration patterns, behavior and the reduction of environmental threats. These studies rely on biochemical analysis and morphological characteristics of each species and individual. Individual identification of sea turtles through manual methods, semi-automatic, photo-identification, of advanced technologies such as artificial intelligence and neural networks are of great relevance in these scientific studies. Photo-identification, geometric morphology of facial scales and dorsal carapace scutes present significant challenges due to their complexity. Natural variability within the same species, changes in morphological characteristics over time, required detailed analysis, and the need for meticulous comparisons complicate the process of individual identification. Current methods, often manual or semi-automatic, require through and sometimes subjective analysis to distinguish between individuals, resulting in a laborious and error-prone process. Additionally, current photo-identification algorithms such as WildID, Hotspotter and I3S-Pattern require images with acceptable quality and where users manually select the rectangular region of interest.In image processing, automatic semantic segmentation techniques precisely delimit specific areas within them. This allows the detection, identification and differentiation of objects through the analysis of meticulous mathematical methods at the pixel level and visual patterns, highlighting contours and specific characteristics of objects in the image content. In this work we propose an artificial intelligence and deep learning model for individual identification of chelonians through multiple convolutional neural networks and mathematical algorithms, with a fundamental role in diverse tasks, ranging from automating semantic segmentation of regions of interest in sea turtle images to edge detection and multimodal understanding of text with natural language models and images. Individual identification is based on geometric morphology of facial scales, shape of carapace scutes and consideration of scales present on flippers and neck. This approach makes possible a significantly more accurate and detailed identification, using specific features of the anatomy of these animals.The semantic segmentation model of content in images SAM (Segment Anything Model), is based on convolutional neural networks. It is distinguished by its ability to detect and isolate a wide variety of objects, regardless of their shape, size or context within the image. Given SAM's versatility, we integrate scenarios from more specialized areas, such as the identification of sea turtles.In our implementation, the results reflect high performance in accuracy. These data are supported by metrics evaluation, such as accuracy and sensitivity, providing a quantitative measurement of the model's ability to identify individuals.Individual identification of sea turtles has a significant impact by providing detailed data on the status and health of populations. This allows it possible to identify and protect critical areas, nesting areas and migratory routes for the balance and health of marine ecosystems.

The introduction of Fastloc-GPS technology to marine animal tracking in the 2010s revolutionised the study of spatial ecology and improved our knowledge of home range sizes. The technology enabled an increased volume of high-accuracy GPS locations, revealing that home ranges of marine vertebrates such as sea turtles are smaller than previously thought in certain populations. Fastloc-GPS data have contributed to quantifying habitat use, understanding the role of environmental variables like oceanic currents in sea turtle movements, and estimating nesting populations when monitoring was not feasible etc. Despite the quality and quantity of its data, Fastloc-GPS tags are seldom employed due to their high costs. Instead, Argos satellite transmitters are the most widely used tool for tracking sea turtles. Considering the low accuracy associated with Argos locations, it is important to understand how these data translate into home range estimates of sea turtles. To test the potential impact of tracking methods on habitat use estimates, we conducted a meta-analysis of green turtle (Chelonia mydas) and hawksbill turtle (Eretmochelys imbricata) home ranges at their foraging grounds (1983-2024). We also generated simulated locations resembling those produced by Fastloc-GPS and Argos tracking technologies to check for variation in estimates due to tracking methods. Further, we explored differences in estimates due to analytical approaches such as Minimum Convex Polygon (MCP), kernel density estimation using package adehabitatHR, Biased Random Bridge and dynamic Brownian Bridge Movement Model. A one-way ANOVA comparison of foraging home ranges of hawksbill and green turtles from 91 studies showed that estimates recorded by Argos differed significantly from other tracking methods including Fastloc-GPS (F=44.76, df=3, p<0.05). There was no significant difference in home ranges as MCPs or utilisation distributions (F=1.78, df=1, p=0.19). Similarly, there were significant differences in 50% and 95% isopleths of home ranges for simulated data of tracking methods i.e., simulated low-accuracy and high-accuracy Argos data, and both Argos data and simulated Fastloc-GPS data (50%: F=114.7, df=2, p<0.05; 95%: F=102.3, df=2, p<0.05) but not between estimates calculated using the four analytical approaches (50%: F=0.06, df=3, p=0.97; 95%: F=0.08, df=3, p=0.97). Based on published studies and simulated data analysis, sea turtle home ranges could sometimes be massively overestimated (by >100-times) using Argos data compared to Fastloc-GPS data. Methodological bias in home range estimation can be attributed mainly to tracking methods and a lesser extent, analytical approaches. Argos locations may overestimate home ranges and could skew our understanding of sea turtle space use and behaviour. While the costs associated with Fastloc-GPS tags are a constraint, their multi-fold application and accuracy are crucial in implementing practical spatial conservation strategies. Fastloc-GPS data provide reliable measures of animal space use that can be used to assign regions of importance for species protection and serve as indicators of habitat quality to inform ecosystem-level conservation efforts. Its application can extend to testing the efficacy of existing Marine Protected Areas boundaries in protecting coastal and marine habitats. Thus, the use of Fastloc-GPS technology could improve designing, monitoring and maintenance of very large Marine Protected Areas or other area-based solutions.

Classification and identification of segmented images into superpixels using Graph Convolutional Networks (GCNs) is an emerging and useful research area. These identification algorithms represent the image content in superpixels, which are the nodes, thus generating a characteristic graph for each image. The different nodes are linked based on similar characteristics. Using GCN with superpixels, better identification can be achieved by transforming complex image information into graph signals to apply deep learning algorithms.

Manual monitoring of sea turtles is a practice that allows for the collection of precise and detailed data. requiring highly trained and experienced human observers to accurately identify different species of sea turtles. Algorithms such as HotSpotter have been developed for animal species identification through photo-identification. These technological tools provide valuable information for research evaluating aspects such as geographic distribution, population density, migration, survival, reproduction, and growth of these species. Additionally, they are quite useful for monitoring the success of sea turtle conservation and protection programs. The aim of this work is to present a solution to address the limitations of manual monitoring through automated identification of individual sea turtles, using their morphological uniqueness and natural patterns such as head structure, shell coloration, fins, and the patterns of characteristic shields of each species in GCNs.

The deep learning algorithm is based on the generation of graph structures, where each edge represents the distance and the color intensity relationship with the different color space channels. Node generation in the graph structure must be coherent with the image content to obtain information in algorithm training, emphasizing the most unique and inherent characteristics of each individual. The GCN algorithm is considered efficient in terms of processing operations compared to other methods. This efficiency stems from processing only 150 superpixels instead of all pixels in the images. The proposed algorithm is robust for images taken in marine and terrestrial environments and under different angles and lighting conditions.

The proposed GCN for sea turtle identification presents high performance in terms of precision and accuracy. Additionally, it offers a significant improvement in sea turtle identification compared to other algorithms, such as Hotspotter. This advancement is achieved by using segmentation to simplify images and leveraging the turtle morphology to generate the graph. Additionally, the use of different color channels for the graph allows for the use of images in various environments, thus eliminating the need for images with specific characteristics for training GCN. We present a new algorithm as a robust and versatile tool to support the monitoring and protection of these species, thus contributing to the understanding of marine life.

To develop appropriate conservation and management methods for endangered sea turtles, it is necessary to efficiently monitor abundance and size composition of foraging aggregations. Unmanned aerial vehicles (UAVs) have recently been used for this purpose, but detecting green turtles foraging in coastal seagrass beds using an automatic object recognition algorithm has been difficult owing to the relatively high false positive rate in the complex natural environment of mixed seagrass beds and coral reefs. Therefore, this study aims to develop a deep learning model to automatically detect green turtles foraging in coastal areas and estimate abundance and size composition from UAV video footage. We constructed a dataset (n = 103,308) of green turtles from UAV video footage taken on Ishigaki Island and Chichijima Island, Japan. The automatic detection model was trained based on the YOLOv7 network. Then, Multi Object Tracking (MOT) was implemented to track each of the green turtles in the video and to assign ID numbers. After threshold filtering by the minimum number of consecutive detections of individuals, the IDs were counted. The automatic detection model based on YOLOv7 resulted in a precision of 0.850 and a recall of 0.859, and mAP@0.5 (mean average precision for the intersection-over-union threshold of 0.5) of 0.922. The model achieved the lowest frequency of false positives when compared to previous studies. Transfer learning did not improve the model performance, and increasing classification classes for status or behavior of green turtles decreased the model performance. Nonetheless, relatively high accuracy in this automatic detection model based on YOLOv7 will help to automatically detect green turtles in natural environments from UAV video footage. We are also developing automatic body size estimation of green turtles.

Photo-Identification (ID) allows to individually identify animals using their unique marks. However, the challenge of ID arise when dealing with large datasets, as manually processing the images can be very time-consuming. Leveraging the advancements in Artificial Intelligence (AI) methods, such as machine learning (ML) algorithms, particularly the deep learning (DL), facilitates automatic processing of image content, and results in higher accuracy in identifying each individual. In this paper, a self-developed algorithm based on ML and DL methods for ID is presented, we compare this algorithm with the HotSpotter ID algorithm for individually identifying sea turtles within populations. Both algorithms employ classical matching methods but differ in their image pre-processing and processing techniques, which involve feature extraction, as well as their approach to One-vs-Many matching evaluation. This paper analyses the characteristics of both methods used for morphological pattern comparison, evaluating the performance of these algorithms within the context of a shared database of sea turtles. The database includes images with acquired content information with various characteristics, i.e., these images present content in different environments, in different light conditions, with a variety of qualities and resolutions, at varying distances, and in different turtle positions and angles. For this comparison, raw images are employed. These unprocessed images offer significant flexibility for post-processing and analysis. Therefore, regardless of how the algorithms operate, the input images they receive do not impact any of the processes conducted by the algorithms. Our approach involves of automatically detecting the head of the turtle with the modified YOLOv5 convolutional neural network, extraction of pixels with high variation to emphasize head scale patterns and utilizing the ORB algorithm for individual identification. The in-house developed algorithm requires no manual intervention for dataset preprocessing. Using fewer computational resources, it allows for the highly accurate identification of individuals in images of standard quality. On the other way, HotSpotter works by manually defining a Region of interest (ROI), indicating by the user the orientation of each image, additionally it also requires manual intervention in selecting the query image that will be compare with the rest of the others. Afterward, it involves extracting distinctive features from images, to identify the closest match. Our proposed photo-identification algorithm, using non-invasive techniques, provides with standard quality images an automatic identification of individual sea turtles with 98% accuracy, whereas the HotSpotter algorithm achieves 96% accuracy. These algorithms are less accurate with images to which motion blur has been added. AI and computer vision tools can be applied to assist in tracking and monitoring wildlife, supporting researchers in conservation efforts. By helping the study of sea turtle populations through non-invasive techniques for individual identification. By simplifying the identification process through AI and computer vision tools, it reduces the need for manual intervention, providing valuable support to researchers and conservationist in their efforts to track, monitor and protect these endangered species.

In the last two decades, some green turtle populations around the world have recovered to high densities. In Lakshadweep, India, systematic studies have documented the loss in productivity and shifts in seagrass species composition in response to green turtle herbivory. Our long-term monitoring programme spanning a little over a decade found that green turtle densities were positively correlated with seagrass shoot densities. A decline in green turtle abundance recorded in each lagoon coincided with low seagrass shoot densities post a period of heavy grazing. This implies that green turtles exhausted forage resources in one lagoon before moving to the next.

Evidently, movements of these turtles en-masse to new foraging grounds can cause trophic cascades and modify entire seagrass ecosystems. Hence, environmental drivers that trigger the behavioural ecology of green turtle movements need to be examined. In collaboration with Arcturus Inc, an indigenous tech company, we supported the development of a low-cost LoRa-based GPS telemetry system at 10-20% of the cost of commercially available tags. Inexpensive radio tags confer a numerical advantage to sample sizes that most research on animal movements lack. With a network of receivers across the lagoons of Lakshadweep, we plan to track multiple green turtle movements to explore their interactions in a larger seascape. The data generated by this study can provide insights into the process and mechanism of establishing new foraging sites, which will prove to be vital for the conservation and management of this species. We aim to make this technology open source to benefit researchers and organizations working in similar systems but particularly in the global south. We have also initiated a non-intrusive photo-identification citizen science study to create a repository of individual green turtles inhabiting the archipelago. In addition to engaging the local community, the movement of individual green turtles will be monitored over the long term.

Photo identification involves classifying unique features of a specific individual. The distinguishing feature used in most sea turtle photo ID studies are the scale patterns on the head. Yet the scale patterns on the turtles' flippers are arguably more complex and could provide an alternative and more robust area for photo ID. Here, we compared the accuracy of the Automatic Photo Identification Suite (APHIS) software to identify individual juvenile and subadult green turtles (Chelonia mydas) based on scale patterns on either the head or the flippers. Photographs were taken using standardized guidelines and then analyzed via APHIS after manually placing marks at intersection points between all scales around a predefined area. We tested whether using 6, 10, or 14 scales influenced accuracy of identifications, and determined that incorporating 14 scales provided the most correct identifications (1st rank) for both head and flipper photo ID. After determining the most accurate location for identification for the head and flippers (dorsal view of the head and digits of the fore-flipper), we conclude that photo ID using flipper scales in APHIS can identify individuals with higher accuracy (100%) than head scales (86%). Nevertheless, as turtles may contort the shape of their flippers during natural movements while the surface of the head remains rigid, photo ID for flippers may currently only be suitable when the flipper can be maintained in a flat position.